Even if CONFIG_COMPAT_BRK is set in the kernel configuration, it can still 
be overriden by randomize_va_space sysctl.

If this is the case, the min_brk computation in sys_brk() implementation 
is wrong, as it solely takes into account COMPAT_BRK setting, assuming 
that brk start is not randomized. But that might not be the case if 
randomize_va_space sysctl has been set to '2' at the time the binary has 
been loaded from disk.

In such case, the check has to be done in a same way as in 
!CONFIG_COMPAT_BRK case.

Signed-off-by: Jiri Kosina <jkosina@suse.cz>
---
 mm/mmap.c |   10 +++++++++-
 1 files changed, 9 insertions(+), 1 deletions(-)

diff --git a/mm/mmap.c b/mm/mmap.c
index 50a4aa0..35d9f9c 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -253,7 +253,15 @@ SYSCALL_DEFINE1(brk, unsigned long, brk)
 	down_write(&mm->mmap_sem);
 
 #ifdef CONFIG_COMPAT_BRK
-	min_brk = mm->end_code;
+	/*
+	 * CONFIG_COMPAT_BRK can still be overridden by setting
+	 * randomize_va_space to 2, which will still make mm->start_brk
+	 * to be arbitrarily shifted
+	 */
+	if (mm->start_brk > mm->end_code)
+		min_brk = mm->start_brk;
+	else
+		min_brk = mm->end_code;
 #else
 	min_brk = mm->start_brk;
 #endif
-- 
1.7.3.1

--

[ message continues ]

Hmm, actually no, that will bring us back into pre-a5b4592c era, please 
disregard that. 

I will have to come up with something more clever (we can't just check 
'randomize_va_space == 2', as that would be racy for already running 
programs).

-- 
Jiri Kosina
SUSE Labs, Novell Inc.
--

[ message continues ]

The patch below handles all the cases properly. Ingo, Andrew, please 
consider applying. Thanks.



From: Jiri Kosina <jkosina@suse.cz>
Subject: [PATCH] brk: fix min_brk lower bound computation for COMPAT_BRK

Even if CONFIG_COMPAT_BRK is set in the kernel configuration, it can still
be overriden by randomize_va_space sysctl.

If this is the case, the min_brk computation in sys_brk() implementation
is wrong, as it solely takes into account COMPAT_BRK setting, assuming
that brk start is not randomized. But that might not be the case if
randomize_va_space sysctl has been set to '2' at the time the binary has
been loaded from disk.

In such case, the check has to be done in a same way as in 
!CONFIG_COMPAT_BRK case.

In addition to that, the check for the COMPAT_BRK case introduced back in 
a5b4592c ("brk: make sys_brk() honor COMPAT_BRK when computing lower 
bound") is slightly wrong -- the lower bound shouldn't be mm->end_code, 
but mm->end_data instead, as that's where the legacy applications expect 
brk section to start (i.e. immediately after last global variable).

Signed-off-by: Jiri Kosina <jkosina@suse.cz>
---
 mm/mmap.c |   10 +++++++++-
 1 files changed, 9 insertions(+), 1 deletions(-)

diff --git a/mm/mmap.c b/mm/mmap.c
index 50a4aa0..ca2f164 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -253,7 +253,15 @@ SYSCALL_DEFINE1(brk, unsigned long, brk)
 	down_write(&mm->mmap_sem);
 
 #ifdef CONFIG_COMPAT_BRK
-	min_brk = mm->end_code;
+	/*
+	 * CONFIG_COMPAT_BRK can still be overridden by setting
+	 * randomize_va_space to 2, which will still make mm->start_brk
+	 * to be arbitrarily shifted
+	 */
+	if (mm->start_brk > PAGE_ALIGN(mm->end_data))
+		min_brk = mm->start_brk;
+	else
+		min_brk = mm->end_data;
 #else
 	min_brk = mm->start_brk;
 #endif
-- 
1.7.3.1


--

[ message continues ]